Working Notes

Why memory. Every other panel here watches demand for compute, or the financing of it. Memory is the input. One projection puts DRAM and NAND at 47% of major cloud capex in 2026 and 68% in 2027 [R84] — a projection resting on extraordinary pricing assumptions, not an observation. If it is even directionally right, memory is not a component of the capex story. It is most of it.

Margin is the symptom. Inventory is the instrument. Micron's gross margin went from 36.8% to 84.6% over six quarters while days of inventory fell from 161 to 122. Selling faster and pricing higher at once is what an unrelieved shortage looks like. The turn appears here first: days rising while margin is still high, because a warehouse fills before a management team changes its language.

Two limits. SK Hynix and Samsung file in Korea and have no XBRL path, so the two largest HBM suppliers cannot appear in the table at all. And EDGAR lags the press release by 22–30 days — the value of the series is the trend, not the freshness.

HBM takes roughly three times the wafer volume of standard DDR5 for the same capacity [R89]. A maker shifting capacity to HBM uses more of the fab to produce fewer total bits, tightening conventional DRAM at the same time. One allocation decision, two markets.

Not yet instrumented: HBM contract pricing and order pushouts, TSMC CoWoS lead times, forward capex guidance.

Two prices for the same inference. Meta's Muse API charges $1.25/$4.25 per million input/output tokens if your prompts stay private, and $0.10/$0.20 if Meta may train on them [R91]. On a 30:1 agentic mix the gap is about $1.24 per million, so a billion tokens a day costs roughly $453,000 a year in forgone discount to keep your data out of training [R92].

The spread is a published valuation of interaction data, not a discount schedule. Where inference is sold below cost to buy data, the list price is two transactions averaged together — and the cost-per-task and frontier-capability panels both assume the list price is revenue per token.

OpenAI's Astra launched the same week at a reported $10/$50 per million [R90]. The prices are moving apart, not together.

Every other instrument on this page watches demand for compute or the financing of it. None watched the input that has become the largest single line in the bill. One projection puts DRAM and NAND at 47% of major cloud providers' capital expenditure in 2026 and 68% in 2027R84. If that is even directionally right, memory is not a component of the capex story. It is most of it.

Filed figures, not estimates — derived from SEC XBRL for the six quarters through each issuer's latest filed periodR85. SK Hynix and Samsung file in Korea and have no equivalent path, so the two largest HBM suppliers cannot appear in this table at all.

Days of inventory is the instrument, not the margin. Margin tells you scarcity is being priced; inventory tells you whether it is ending. It comes from two tagged concepts with no judgement in between, and memory downturns appear in warehouses before they appear in management language. Micron's fell from 161 days to 122 while its gross margin went from 36.8% to 84.6% — it is selling faster and pricing higher at the same time, which is what an unrelieved shortage looks like. The turn, when it comes, shows up here first: days rising while margin is still high.

Reported capex, not guidanceR86. Guidance never appears in a filing, which is the same measurement gap this monitor's balance-sheet instruments already carry. Against this denominator, the memory share above is the numerator nobody is filing.

SK Hynix held about 58% of HBM revenue in the first quarter of 2026, with Samsung and Micron near 21% eachR87. In the same window Samsung led total DRAM revenue, and by the second quarter held about 39% against SK Hynix 26%, Micron 25% and CXMT 7%R88. Both are true. “Who is winning memory” has two opposite answers depending on what is counted — which is why every share figure on this page names its denominator and its period.

The structural fact underneath: HBM consumes roughly three times the wafer volume of standard DDR5 for the same capacityR89. A maker shifting capacity to HBM is using more of the fab to produce fewer total bits, which tightens conventional DRAM at the same time. One allocation decision moves two markets, and it raises the cost of everything else in a datacentre, not just the accelerator.

Two launches in the same week moved list prices in opposite directions. Read together they break a measurement this monitor depends on: one of the two prices is not being paid in cash.

Priced at a reported $10 per million input tokens and $50 per million outputR90. The detail that matters here is not the price but the architecture: OpenAI describes capability increasingly expressed with fewer or no language tokens, which reduces what chain-of-thought monitoring can observe. That does not make verification impossible — it moves what can be checked from the reasoning to the outcome, which is a narrower instrument.

92% off input, 95% off output for the same inferenceR91. The spread is not a discount schedule, it is a published valuation: the gap is roughly $1.24 per million tokens, which is what Meta will pay for interaction data. An enterprise running a billion tokens a day pays about $453,000 a year to keep its data privateR92.

Why this sits beside the memory panel. They are the same measurement problem from opposite ends. Memory share of capex asks what an operator pays to produce a token. The Contributor tier answers what a token is worth to the producer — and the answer is that inference can be run at a loss to acquire training data. Where that is happening, the headline token price is two different transactions averaged together, and any elasticity or unit-economics read that treats it as revenue per token is measuring neither. This monitor's cost-per-task and frontier-capability instruments both assume the list price is revenue. That assumption now has a boundary.

The buildout is justified by the premium that the most capable models command. So it is worth asking what that premium actually is. Each point below is the cheapest model available at or above its capability level — the efficient frontier of the market as it is priced today. Cost rises with capability throughout, and then accelerates sharply near the top: the steepest single step on the curve buys the last stretch of measured capability.

The capability/cost frontier, September 2026. Capability is the Epoch Capabilities Index; cost is list price per million tokens, blended 3:1 input:output, on a logarithmic axis. Open-weight models are marked separately.R76

Read the right-hand column. Three quarters of the entire capability range costs under ten cents per million tokens — 1.9x, 1.9x, 1.7x across the whole of it. Then the curve does not bend, it breaks: 106x across the next fifteen percentiles. In absolute terms, index 154.5 costs $0.094 and index 162.6 costs $10.00: a 5.2% gain in measured capability costs a hundred and six times as much.

Revised 9 September 2026. Two things moved at once and they should not be confused. Epoch refits its index as benchmarks are added and retired, so every score on this panel changed together — a scale change, not a correction. Separately, the price basis changed: the panel now uses the cheapest place a model can actually be bought, and a vendor's own list price only where that vendor is the only source. The earlier version of this note read “1.9x, then 3.5x, then 4.5x… the 75th-to-90th band costs 10x.” The break is in the same place and is far sharper than published.

The composition of the frontier splits just as sharply, and along the same line. All five frontier models below the break at index 156 are open-weight or Chinese. All five above it are closed and American. Not a tendency — every single one. The cheap frontier is largely open; the expensive frontier is entirely closed and American. That is consistent with the separate finding that seven of the ten most-used models on OpenRouter are open-weightR56.

One of these prices has an expiry date printed on it. Gemini 3.7 Flash sits on the frontier at $1.50 blended, but Google's own pricing page states that rate holds only through 31 December 2026, rising to double on 1 January 2027R77. At the new rate it leaves the frontier entirely. A frontier partly composed of introductory pricing is not a stable frontier — and an aggregated price feed reports the current number with no expiry attached. This one was visible only by reading the vendor's page.

Capability: Epoch AI, "AI Benchmarking Hub", published online at epoch.ai, licensed CC-BY. Cost: each vendor's own published pricing page, read 20 August 2026, with an MIT-licensed aggregated price map used as a cross-check. Where the two disagreed by more than 5% the vendor page was used — this happened once, on GPT-5 nano, where the aggregator was 9.1% high.

The cost axis is price per token, not price per task. A verbose reasoning model emits more tokens to answer the same question, so per-token price understates its true cost per task — and understates it most at the top of the range, which is exactly where this chart is most interesting. The direction of that bias is known; its size is not.

Capability indices are constructions, not measurements. The index used here is a composite over one organisation's benchmark suite. A different suite would move the points; the question is whether it would move the shape.

200 of 283 indexed models carry a published price. Every current-generation unpriced model in the band where it could displace a frontier point was priced by hand against its vendor's page, because a missing price silently drops a model out of the frontier — and dropping a cheap capable model would exaggerate the premium in the direction this page already argues. The remainder are superseded previews and duplicate host listings.

Two different kinds of price sit on this chart. Where a model's creator sells it, the creator's published list price is used. Where the creator only released weights and runs no API of its own — gpt-oss, Gemma, Mistral NeMo, Qwen2.5-Coder — there is no such price, so the cheapest third-party hosted rate is used instead. Hosted rates for a single model span up to 19x across providers; the minimum is taken deliberately, because the frontier asks what is cheapest available.

The capability index does not cover every lab. Tencent, among others, has no entry in it at all, so its models cannot appear here however cheap or capable they are. A benchmark suite that omits a vendor removes that vendor's points from the cheap end of the frontier, which flatters the premium rather than understating it.

This prices cloud inference, and only cloud inference. Every figure here is what somebody charges to run a model on their machines. Work that migrates onto the buyer’s own hardware — an open-weight model on a workstation, or in time a phone — leaves this series entirely, and its departure looks identical on the chart to a fall in the cost of production. Gavin Baker, asked for the case against the buildout, volunteered exactly this: edge AI is “by far the most plausible and scariest bear case”. We would measure a deflation and report a cost curve, and nothing in this panel would distinguish the two.

List prices only. No batch discount, no cache discount, no negotiated enterprise rate. Large buyers do not pay these numbers.

• The cost axis is price per token, not price per task. A verbose reasoning model emits more tokens to answer the same question, so per-token price understates its true cost per task — and understates it most at the top of the range, which is exactly where this chart is most interesting. more →
• List prices only. No batch discount, no cache discount, no negotiated enterprise rate. more →
• Capability indices are constructions, not measurements. The index used here is a composite over one organisation's benchmark suite. more →
• 200 of 283 indexed models carry a published price. Every current-generation unpriced model in the band where it could displace a frontier point was priced by hand against its vendor's page, because a missing price silently drops a model out of the frontier — and dropping a cheap capable model would exaggerate the premium in the direction this page already argues. more →
• Two different kinds of price sit on this chart. Where a model's creator sells it, the creator's published list price is used. more →
• The capability index does not cover every lab. Tencent, among others, has no entry in it at all, so its models cannot appear here however cheap or capable they are. more →
• This prices cloud inference, and only cloud inference. Every figure here is what somebody charges to run a model on their machines. more →

Both charts above price capability per token, which is the number vendors publish and the wrong number for anyone buying work. This one uses measured spend: Cursor runs a benchmark of real, ambiguous, multi-file coding tasks and reports what each model actually cost to finish one. No blended rate, no assumptions — just the bill.

The cheapest configuration reaching each score, against measured cost per task. Each model appears several times — once per reasoning-effort setting.R80

The shape of the first chart survives on measured money. Going from 70.8% to 72.9% — just over two points of score — costs 6.4x, from $2.81 to $18.02 a task. That is the same conclusion the capability/cost frontier reached from list prices and an abstract index, arrived at independently with no pricing assumptions at all. Two constructions agreeing is worth more than either alone.

It also shows something list prices cannot. The effort dial is the cost dial. Claude Fable 5 costs $6.80 a task at medium effort and $18.02 at maximum — a 2.6x range for one model at one published price, which is more than switching vendor buys across most of the range. The panels above price a model. This one prices a decision about how hard to run it.

Which means the per-token charts above are not understating the cost of this work — if anything they overstate the rate, because agentic tasks are dominated by input tokens and input is a fifth the price of output. The cost comes from volume, not from the rate. So there is no correction factor: a per-token price simply cannot tell you what a task costs, in either direction. That is why the charts above say what they measure rather than adjusting for it.

Source: CursorBench, cursor.com/cursorbench, and ARC-AGI-2, both via Epoch AI, "AI Benchmarking Hub", epoch.ai (licensed CC-BY). Read 21 August 2026. List prices used for the reconstruction are the same vendor pages as the panels above; the cross-domain test uses no price data.

Measured: the score, the cost per task, the steps per task, and the token count Cursor reports. Cursor states cost is computed by applying each model's published pricing for input, cache read, cache write and output to the tokens it used.

Inferred: the million-token figure. Cursor does not define what its token column comprises; the arithmetic implies it is output only. Because cache reads bill at about a tenth of input, the reconstruction charges those tokens too much, so a million and a half is a floor, not an estimate — the true volume is higher. Both corrections push the same way, which is why the conclusion holds even though the field is undocumented.

Tested against a second domain, and it held. An earlier version of this note said nothing here generalises. That was too strong. ARC-AGI-2 is abstract visual reasoning — no repository, no tool loop, no resent context — and 32 model configurations appear on both benchmarks with the identical effort setting. For those, the price per token is the same number on both sides and cancels exactly, so the ratio of costs is the ratio of token volume, with no pricing assumption at all. An agentic coding task costs a single-digit multiple of an abstract-reasoning task — median 1.72x against ARC-AGI-2 and 4.20x against the easier ARC-AGI-1 — and on the harder benchmark 7 of the 32 are cheaper on the coding side. Million-token tasks are not peculiar to agent loops. Quote the range, not one figure: the estimate moves with the difficulty of whichever reasoning benchmark sits in the denominator, because an easier one burns fewer tokens. The order of magnitude replicates; the point estimate does not.R82

Two domains is not all domains. Both benchmarks tested here are long-generation work. Chat, summarisation, classification and retrieval remain untested, and a single-turn task should be smaller by orders of magnitude. The caveat has narrowed, not disappeared.

These are the benchmark's costs, not a buyer's. No enterprise discount, no prompt engineering to cut steps. A competent operator pays less.

Coverage is 67 configurations across four model families. A cheaper model outside those families would not appear, and the 6.4x step at the top is a jump between two specific models rather than a smooth market curve.

This prices cloud inference, and only cloud inference. Every figure here is what somebody charges to run a model on their machines. Work that migrates onto the buyer’s own hardware — an open-weight model on a workstation, or in time a phone — leaves this series entirely, and its departure looks identical on the chart to a fall in the cost of production. Gavin Baker, asked for the case against the buildout, volunteered exactly this: edge AI is “by far the most plausible and scariest bear case”. We would measure a deflation and report a cost curve, and nothing in this panel would distinguish the two.

• Measured: the score, the cost per task, the steps per task, and the token count Cursor reports. more →
• Inferred: the million-token figure. Cursor does not define what its token column comprises; the arithmetic implies it is output only. more →
• Tested against a second domain, and it held. An earlier version of this note said nothing here generalises. more →
• Two domains is not all domains. Both benchmarks tested here are long-generation work. more →
• These are the benchmark's costs, not a buyer's. No enterprise discount, no prompt engineering to cut steps. more →
• Coverage is 67 configurations across four model families. more →
• This prices cloud inference, and only cloud inference. Every figure here is what somebody charges to run a model on their machines. more →

The panel above prices capability against an index. This one prices it against something a buyer actually decides about: how long a task a model can carry out on its own. METR measures that directly — the task duration, in human time, at which a model succeeds half the time. Plotted against the same list prices, it moves the story. The expensive step is not at the top. It is the step into hour-long work.

The cheapest model able to sustain a task of each length, against blended list price per million tokens. Both axes logarithmic. Bars show METR's fitted interval, which is wide at the top.R78

Crossing from half-hour to hour-long tasks costs 19.2x. Going from one hour to eight costs about five times in total. Once a model can sustain an hour of autonomous work, extending that horizon is comparatively cheap. Acquiring the first hour is what costs. That is the opposite shape to the index panel above, and both are true — they are different axes. But they answer different questions, and this is the one a buyer faces.

Now the part that matters more than the price curve. Everything above is measured at a 50% success rate — a coin flip. METR also publishes the horizon at 80%, which is nearer a bar anyone would actually deploy against. The horizons do not shrink gently. They collapse.R79

The twelve-hour model is a seventy-minute model at a threshold you would rely on. And the ranking inverts: at 80%, the most expensive model on the chart drops off the frontier entirely, beaten on horizon by one costing less than half as much. The buildout is underwritten by an expectation of long autonomous capability. Measured at a bar anyone would deploy against, the best horizon in this data is about ninety minutes — and paying twice as much does not buy more of it.

Capability: METR, "Measuring AI Ability to Complete Long Tasks" (arXiv:2503.14499) and "Task-Completion Time Horizons of Frontier AI Models" (Time Horizon 1.1), metr.org/time-horizons. Horizons are the 50% success threshold except where the 80% figure is named. Cost: vendor list prices, read 20 August 2026, on the same basis as the panel above. The two models on this frontier that are still sold were checked against their vendors' pages and both matched exactly. METR measured the horizons. The frontier construction, the pricing and every conclusion drawn on this page are this monitor's own, and METR does not underwrite any of them. METR confirmed on 21 August 2026 that its public work may be cited on that basis.

The uncertainty is large, and largest where the chart is most interesting. Horizons are fitted curves, not stopwatch readings. Claude Opus 4.6's twelve hours carries a fitted range of 5.3 to 65.8 hours, which overlaps Gemini's substantially. The shape — a steep step at the bottom, a flat stretch above it — survives that. The top-end level does not, and nothing here should rest on it.

This is price per token, not price per task, and the mismatch is sharper here than on the panel above: an eight-hour agentic job emits vastly more tokens than a thirty-minute answer. The real spread across task lengths is therefore wider than shown. This is the price of the tokens, arranged by task length.

Three of the six models on this frontier are no longer sold. OpenAI has withdrawn o4-mini and GPT-5 from its price list; Moonshot has withdrawn K2 Thinking. Those figures are historical. This is structural, not sloppy: measuring long-horizon capability takes long enough that the models measured have since been retired. Read this as a snapshot of a market that has already moved.

• The uncertainty is large, and largest where the chart is most interesting. Horizons are fitted curves, not stopwatch readings. more →
• This is price per token, not price per task, and the mismatch is sharper here than on the panel above: an eight-hour agentic job emits vastly more tokens than a thirty-minute answer. more →
• The newest models are absent. Claude Opus 5, Fable 5 and the GPT-5.6 family have no measured horizon at all — and they are the expensive ones, so this understates what the current top tier costs.
• Three of the six models on this frontier are no longer sold. OpenAI has withdrawn o4-mini and GPT-5 from its price list; Moonshot has withdrawn K2 Thinking. more →

The claim is that chip vendors and hyperscalers fund their own customers, so some reported revenue is the seller's own money returning. A register of 29 announced arrangements was built and every quotable figure traced to an SEC document or withdrawn. The exercise refuted more of the thesis than it confirmed — eight refuting findings against twelve confirming onesR30b.

The structures are built to stay non-voting and below the significant-influence threshold, so that ASC 850 related-party disclosure never fires. At Microsoft it fired. The October 2025 recapitalisation pushed its holding to an ~25% as-converted equity-method interest, which compelled the exact figure the structure was meant to keep private: $24.1B of FY2026 revenue from OpenAI, $6.0B receivableR20.

The three peers holding comparable positions disclose none of itR21. So the rule is neither that ASC 850 never fires nor that it does: the avoidance structure holds at three of four filers and fails only when a recapitalisation pushes a stake across the line. That makes Microsoft's $24.1B a natural experiment for what the other three are not required to reportR22.

Oracle's $300B OpenAI contract and AMD's ~$90B could not be traced to any filing by either party. A sweep of 212 documents — every Oracle and AMD accession since mid-2025, exhibits and XBRL included — returns zero hits for $300 billion and zero for Stargate. Both figures were carried here and have been withdrawnR23.

The $0.01 warrant is an industry instrument, not a one-off: AMD→OpenAI, AMD→Meta, Google→TeraWulf, Google→Cipher, CoreWeave→Core Scientific. A warrant struck at one cent is stock delivered on a conditionR25. Alphabet discloses $43.8B of credit derivatives, nearly tripled in six months from $16.9BR26. But Alphabet names no counterparty, so the linkage to those specific deals is inference, not disclosureR27.

The margin test came back clean — AWS margin expanded ~645bps while growth accelerated, Google Cloud margin roughly doubledR28. That is narrower evidence than it appears. The test sees concessions delivered as price and is blind to concessions delivered as equity — for the Nvidia, Amazon and Microsoft structures it has no power by construction, not merely no signal yetR29.

Method. Every figure is quoted verbatim or marked as inference. Disproven findings are retained with their retraction rather than deleted. The conclusions were put to an adversarial panel instructed to attack them; it returned kill, which is what prompted the 212-document sweep. The claims survived, one Oracle overstatement was corrected, and the margin-test scope was narrowedR30b.

The tell this quarter is not the size of the spend — it is that operating cash flow stopped covering it. Two of the big four are free-cash-flow negative in the same quarter, and Microsoft's relief rally came from holding the envelope flat while extending useful livesR39. Aggregate 2026 guidance now sits near $725–800B, a 77% step-up on 2025R60.

Meta is the exception that qualifies the rest. At Amazon, Microsoft and Alphabet the financial leg deteriorated while demand accelerated. Meta is the first name where both moved together — and it has the least contracted revenue to fall back onR43.

Current values against bubble-top thresholds. Hover any metric for what it means. Six instruments were amended in August 2026 — why.

Construction-in-Progress represents capital spent on data centers not yet active. Assets don't depreciate until "placed in service", letting companies park capital here to defer massive depreciation expenses.

The build-out has crossed from internally-funded to externally-funded. Rising long rates lift the hurdle on every incremental capex dollar just as incremental returns compress, and directly squeeze the debt-service coverage of leveraged neo-clouds borrowing above 10%.

For ad-funded capex (Meta, Alphabet), impression volume is the demand signal underneath the revenue line. Revenue held up by price per ad while volume decelerates is the classic late-cycle tell — the engine funding the build-out is weakening beneath a flattering headline.

Q2 2026 break: Alphabet printed 0.87x — its first negative free cash flow as a public company. Meta's coverage left only $784M of FCF.

This monitor originally watched utilisation and lease rates for signs the leveraged intermediaries could not service their debt. Q2 2026 showed both healthy while losses widened anyway — the failure is arriving through interest expense, not demand.

The original instrument assumed an efficiency breakthrough would arrive from outside and damage Nvidia. It arrived from inside and was bought: Nvidia licensed Groq for $20B. Revenue is insulated; obsolescence risk transfers to owners of the prior generation and their creditors.

Unresolved: Nvidia projects LPU racks supporting $45/M tokens against OpenAI's ~$15 — efficiency raising price. That contradicts the efficiency-deflation thesis. Both cannot hold.

High lead times strand capital: Completed facilities can't get power. Equipment sits idle in non-depreciating CIP, deteriorating ROIC.

Q2 2026: CoreWeave revenue +112%, net loss widened to $626M on financing cost. Nebius adjusted EBITDA +$236M against GAAP −$190M.

External funding meets a rising cost of capital. Promoted to primary read — price now leads issuance volume.R57

All four hyperscalers earn a lower return on invested capital than they did at the start of 2024. They are not failing businesses: operating income grew 36% to 79% over the same period. Their capital bases grew 67% to 160%. Computed quarterly from SEC filings; latest data 30 June 2026.

Pretax ROIC, annualised from the quarter.R74 Q4 is absent by construction — 10-Ks report full-year durations. Microsoft’s fiscal year ends in June.

Assets under construction sit in the denominator earning nothing, so a falling return could be timing rather than decline. Dashed lines remove that capital entirely. The gap between solid and dashed is the air gap; the slope is the answer.

Excluding it lifts the level by roughly 10–12 points — but does not flatten the decline. Alphabet’s −6.9pp becomes −7.5pp; Meta’s −14.6pp becomes −15.7pp. Both steeper without it.R74 Alphabet’s operating income rose 60% over this window and its return on capital already in service still fell. Meta reports “construction in progress”; Alphabet reports “assets not yet in service”. Microsoft discloses neither and Amazon annually only, so the test covers two of four.

Alphabet’s purchase commitments and other contractual obligations went from $332.4bn to $811.0bn in one quarter. The disclosure heading, categories and wording are identical across both filings, so this is growth rather than a change in what is disclosed. Short-term commitments grew 1.45×; long-dated commitments grew 3.14×, and 87% of the $478.6bn increase is long-dated. Over the same two quarter-ends, invested capital rose $166bn. Two independent measures, one company, one quarter, the same direction.R73

It does not say returns are inadequate. Microsoft at 40% and Alphabet at 23% pretax remain high. The finding is the trajectory, not the level.

By early 2026, Construction in Progress (CIP) balances have reached unprecedented levels. Under GAAP rules, hardware classified as CIP is exempt from depreciation — creating a temporary shield for operating margins as obsolescence risk accumulates off the P&L.

The Fracking Trap: Like shale oil wells, high-end compute clusters suffer rapid technological decline curves. A GPU cluster purchased today might lose 80% of its competitive economic utility inside 36 months, yet it is capitalized with slow, optimistic depreciation schedules.

Subprime Securitization: Off-balance-sheet SPVs and neo-cloud leasing transfer hardware risk to private credit. As of August 2026 this is no longer an analogy — $500B of GPU-collateralised SPE financing was announced, funded by insurance and retirement capital.R19 Long-term construction leases are paired against volatile, cancelable user compute demands — creating a severe asset-liability mismatch.

Contracted forward obligations that sit outside the balance sheet — purchase and unconditional obligations, plus leases signed but not yet commenced — set against each company's trailing-twelve-month capex. The multiple is how many years of current spending is already contracted. Alphabet 6.8×, Meta 7.0×, Microsoft 4.8×, Amazon 1.5×; $2.35tn against $511bn of TTM capex across the four, or 4.6× in aggregate.R73 Read from filing text, not XBRL: these figures are not tagged and structured-data queries return unrelated, far smaller values. On-balance-sheet lease liabilities are excluded for every company so the four are on one basis. Microsoft's $329.1bn is disclosed directly in its FY2026 lease note as leases “primarily for datacenters” that have not yet commenced. As of Q2 2026 for Alphabet, Meta and Amazon; FY2026 (30 June) for Microsoft.

The transition from speculative expansion to normalization is inevitable. Institutional portfolios should pivot toward industrial "bottleneck beneficiaries" like the turbine oligopoly — GE Vernova now filling 2028–29 slots with some customers pulled into 2030R62 — while reducing exposure to leveraged intermediaries dependent on continuous venture injections.

The forensic question is not how much the hyperscalers spend — it is whether the business still generates the cash to pay for it. This is trailing-twelve-month operating cash flow divided by trailing-twelve-month capital expenditure, straight from SEC filings, quarter by quarter. Below 1.0x, the buildout is being funded from the balance sheet rather than from operations.

Figure 4 — Operating cash flow as a multiple of capital expenditure, trailing twelve months. Every tracked builder has compressed toward the 1.0x line since 2022; Oracle and Amazon have crossed it.R75

Source: SEC EDGAR XBRL company filings. Most quarters are derived by differencing year-to-date cumulative filings, since few issuers tag discrete quarterly cash-flow figures. Fiscal years are not aligned — Microsoft's ends in June, Oracle's in May, the rest in December — so read trajectory rather than like-for-like quarters. Trailing-twelve-month basis is used deliberately: single quarters are dominated by working-capital seasonality (Amazon's Q1 2022 operating cash flow was genuinely negative). Filings lag earnings by 22–30 days, so the final point may predate the most recent press release.

The current expansion mirrors the speculative patterns of the 1998–2000 telecommunications build-out and the 2005–2007 subprime credit expansion. While the Dot-Com era was driven by a demand myth regarding data traffic, the current cycle faces a unique reversal: hyperscalers are bearing the costs while their downstream customers remain heavily unprofitable.